feat: tenant isolation Phase 4 — RLS on all remaining tables #136

Merged
chihlasm merged 9 commits from feat/tenant-isolation-phase-4 into main 2026-04-12 04:35:32 +00:00

9 Commits

Author SHA1 Message Date
chihlasm
52553d62d2 fix(tests): update expectations for RLS-correct behavior
- test_rls_isolation: add pytestmark for module-scoped event loop to fix
  "Future attached to a different loop" with pytest-asyncio 0.23 + asyncpg
  module-scoped fixtures
- test_admin_categories_global: global categories use PLATFORM_ACCOUNT_ID
  not NULL; update stale assertion
- test_permissions_account: with RLS, cross-tenant tree access returns 404
  (invisible) not 403 (forbidden) — update to match actual behavior

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
2026-04-12 03:48:30 +00:00
chihlasm
a48660700a fix: background jobs and lifespan must use BYPASSRLS sessions
All code that runs outside a request context (APScheduler jobs,
lifespan startup) has no app.current_account_id set, so the
app-role session returns 0 rows from every RLS-protected table.

Changed to _admin_session_factory (BYPASSRLS) in:
- knowledge_flywheel_scheduler.py — queries ai_sessions
- psa_retry_scheduler.py — queries psa_post_log
- retention_cleanup.py — queries assistant_chats
- scheduler.py (_fire_maintenance_schedule, _cleanup_expired_ai_conversations)
- main.py (archive_stale_ai_sessions, _process_notification_retries,
  load_all_schedules at startup)

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
2026-04-12 03:44:23 +00:00
chihlasm
3ff886363c fix: use BYPASSRLS session for all auth deps and user-mutation endpoints
Phase 4 enabled RLS on the users table. All code paths that touch users
(or other RLS-protected tables) before require_tenant_context sets
app.current_account_id must use get_admin_db (BYPASSRLS):

- deps.py: get_current_user and get_current_active_user → get_admin_db
- auth.py: all endpoints → get_admin_db (login, register, refresh, etc.
  run before tenant context exists; mutation endpoints also need session
  consistency since current_user is in the admin session)
- accounts.py: transfer_ownership, leave_account, delete_account
  → get_admin_db (these mutate current_user directly)
- onboarding.py: dismiss_onboarding → get_admin_db (same reason)

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
2026-04-12 03:25:18 +00:00
chihlasm
501442e5f0 fix: seed_test_users must use ADMIN_DATABASE_URL after Phase 4 RLS on users
RLS is now enabled on the users table. The seed script was using the
app-role connection (DATABASE_URL) which has no tenant context at seed
time — all SELECTs return 0 rows and INSERTs are blocked by FORCE RLS.

Falls back to DATABASE_URL if ADMIN_DATABASE_URL is not set (local dev
without roles configured).

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
2026-04-12 03:12:46 +00:00
chihlasm
6f53ec06f5 docs: add lessons 107-109 — RLS startup, global tables, tree_shares account_id
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
2026-04-12 02:58:12 +00:00
chihlasm
ec322f7cdf fix: bootstrap service account with BYPASSRLS session 2026-04-12 02:44:36 +00:00
chihlasm
f9248aeaa8 fix: remove platform_steps and template_trees from Phase 4 RLS
Both tables have no account_id column — they are globally readable
by all authenticated users and must not have RLS policies.

Also removes the corresponding test cases that assumed these tables
had account_id-based policies.

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
2026-04-12 01:48:50 +00:00
chihlasm
c6da4ebee5 fix: remove script_categories from Phase 4 RLS — no account_id column
script_categories is a global lookup table (shared across all tenants).
The account_id column belongs to ScriptTemplate in the same model file,
not ScriptCategory. The Python scan matched the file, not the class.

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
2026-04-12 01:32:42 +00:00
chihlasm
64f004a62c feat: tenant isolation Phase 4 — RLS on 31 remaining tables + script_builder fix
Enable RLS on all remaining tenant-scoped tables (31 tables):

Standard policy (tenant sees own rows):
  users, account_invites, account_limit_overrides, account_feature_overrides,
  subscriptions, ai_chat_sessions, ai_conversations, ai_session_steps,
  ai_session_embeddings, ai_suggestions, ai_usage, assistant_chats,
  attachments, copilot_conversations, feedback, file_uploads, fork_points,
  kb_imports, notifications, notification_configs, notification_logs,
  psa_activity_logs, psa_member_mappings, script_builder_sessions,
  script_categories, session_ratings, tree_embeddings, user_folders,
  user_pinned_trees

Platform-visibility policy (own rows OR PLATFORM_ACCOUNT_ID):
  platform_steps, template_trees

Intentionally skipped:
  accounts (IS the root table, no account_id column)
  plan_feature_defaults (platform config, no account_id column)

Also fixes script_builder_service.create_session() which was missing
account_id= on ScriptBuilderSession construction, causing 500s on all
script builder endpoints (pre-existing CI failure).

Adds Phase 4 RLS isolation tests covering: users, script_builder_sessions,
ai_session_steps, notifications, platform_steps, template_trees.

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
2026-04-12 01:25:28 +00:00